An improved Spectral Analysis based Subspace Detection for Hyperspectral Image Classification

Hyperspectral imagery, with its rich spectral information, is indispensable in various remote sensing applications. However, working with hyperspectral data introduces two formidable challenges: the curse of dimensionality and the predicament of class imbalance. The inherent high dimensionality necessitates substantial computational resources, while class imbalance undermines overall classification accuracy due to the scarcity of samples in training for certain classes. To mitigate the curse of dimensionality, this research analyses a few diverse feature extraction techniques, including Principal Component Analysis (PCA), Sparse PCA, segmented PCA, and Linear Discriminant Analysis (LDA). Although these methods can reduce the features from the original dataset, however sometimes depend on certain global statistics such as variance. Also, these methods could not handle the small classes as they could not represent the overall characteristics of the samples. Therefore, the proposed method incorporates the solution to the class imbalance problem as well as reduces the features based on the given classes. The ensuing classification phase employs a kernel Support Vector Machine (KSVM), achieving an accuracy of 88.81% only when the class imbalance exists. Later the class imbalances are resolved through the incorporation of Synthetic Minority Over-sampling Technique (SMOTE). The proposed approach reduces the input features and generates synthetic data for each class based on the proportions original simultaneously rectifying class imbalance while preserving dataset characteristics. Remarkably, this strategy results in an improvement of accuracy to 98.69%. This study underscores the significance of hyperspectral imagery, delineates its inherent challenges, and presents novel solutions that substantially enhance classification performance, making hyperspectral data more accessible and valuable in remote sensing endeavors.